Summary:
"Show-off" robots that compete, dance, and perform in bands are still here, but more common are wheeled machines placed in various scenarios to demonstrate their "working abilities": on assembly lines, in front of washing machines, beside coffee makers—they no longer necessarily need human-like "bipedal" legs or lifelike skin, but they must know how to "do something."
July 17, 2026, Shanghai World Expo Exhibition Center. 100,000 square meters of exhibition space, over 1,100 enterprises, over 3,000 exhibits, more than 300 global debuts. Beyond the numbers, a more noteworthy change is happening.
Last year, Unitree's boxing championship match was swarmed by crowds; back then, robot dancing was a staple, but actually performing tasks was rare. This year, however, you'll find a robot traffic officer in a police uniform on the exhibition roads, robot guides in yellow vests at the entrance, robots serving tea, scurrying around, tidying up homes everywhere. At the very least, they can show off some ping-pong skills or lift dumbbells. This year, people have grown indifferent to robot boxing; the more eye-catching products have become the 3.9 million yuan mech robots.
Walking into the WAIC exhibition hall, carbon-based lifeforms feel a sense of intrusion, as if entering a cyberpunk life plaza.
"Show-off" robots that compete, dance, and perform in bands are still here, but more common are wheeled machines placed in various scenarios to demonstrate their "working abilities": on assembly lines, in front of washing machines, beside coffee makers—they no longer necessarily need human-like "bipedal" legs or lifelike skin, but they must know how to "do something."
"Last year, everyone was showing off parameters; this year, the wind has completely changed, everyone is heads-down rushing towards commercialization," a staff member on-site lamented to Phoenix Net Technology.
From Lab to Mass Production: How Many Hurdles Remain?
This year's conference for the first time elevated embodied intelligence to a core track alongside intelligent computing. The number of embodied intelligence exhibitors surged from over 80 last year to over 200. 208 embodied intelligence terminals and over 300 physical machines were showcased together.
The most obvious change is written on the "badges" of dozens of humanoid robots. Last year they were exhibits in booths; this year they directly became WAIC volunteers—roaming various corners of the venue, undertaking guide and inquiry support duties. This marks the first large-scale deployment of full-size humanoid robots at such a large-scale exhibition, with the technology provider being Agibot.
"In the past, people would ask: Can the robot dance? Can it do a somersault? Now people ask: Can the robot work continuously for 24 hours? Can it repeat an action ten thousand times? Can the function be reused for another standardized product?" said Hu Chunxu, Vice President of Developer Ecosystem at Digua Robot. Mass production is becoming the industry's watershed this year. This is no longer a "muscle-showing" carnival, but a systematic industrial exam from lab to factory, from conversation to task execution.
Approaching various robot booths, one also finds changes happening; evaluations of robots are becoming more complex.
Agibot's exhibited Expedition A3 Ultra won the WAIC "Treasure of the Exhibition Hall," the highest official honor of the exhibition segment, focusing on frontier innovation, industrial application prospects, economic benefits, and application value. This product is equipped with a 360° visual and LiDAR fusion perception system, paired with 700 TOPS computing power.
Wang Cong, CEO of Digua Robot, also told us in a pre-WAIC briefing that upon truly entering factory scenarios, many problems originally thought to be algorithmic turned out to be practical issues of production management and system integration, because the real requirements for generalization and stability only emerge after deployment.
For embodied intelligence to move from the lab to mass production, the entire industry must face at least three major hurdles.
The first is the difficulty of model deployment. Parameters of VLA, VLM, and other large models are getting larger; stuffing them onto the edge side requires overcoming hurdles like computing power, memory, quantization, and toolchain adaptation; putting them all in the cloud can't avoid network latency, making real-time robot perception and control impossible to keep up.
The second is high system computing power requirements. A robot isn't done after running one model; data from multiple cameras, LiDAR, and IMUs need spatiotemporal alignment, dozens of joints require low-latency coordination, and perception, reasoning, and control must run simultaneously, demanding far more from the computing architecture than ordinary smart devices.
The third, most easily overlooked, is the difficulty of industrial collaboration. Making a demo in the lab requires success once out of ten tries; manual parameter tuning and temporary wiring can suffice. For true mass production, structural design, interface standards, reliability verification, automated calibration, full-process quality control—none can be missing, impossible for a single enterprise to handle alone.
"In the past, it was about breakthroughs in single-point technology; now it's about the speed of industrial chain collaboration," Digua Robot's relevant person in charge repeatedly mentioned this phrase during the briefing.
The 'Brain' is the Real Bottleneck
Hardware is advancing, but progress on the "brain" is noticeably slower.
Zhu Xing, CEO of AntLingBo, recently told the media: "Over the past year or two, progress in (robot) 'cerebellum' and hardware has been very fast. Compared to the progress of the cerebellum, progress of the brain is slow."
Gu Jie, founder and CEO of Fourier, holds a similar judgment: "By 2026, embodied robot technology architecture has shown a clear convergence trend." But what's converging is the hardware architecture; the brain remains the most uncertain variable.
On ZhiPingFang's booth, the brain-inspired large model NeuroVLA was placed in the most prominent position. This model draws on human brain working mechanisms, adopting a three-layer collaborative architecture of "Cortex—Cerebellum—Spinal Cord"—the cortex handles semantic understanding and task planning, the cerebellum handles high-frequency motor coordination and dynamic correction, and the spinal cord handles millisecond-level motion execution and safety reflexes.
Morgan Stanley has listed ZhiPingFang as a representative enterprise in the robot "brain" direction. Even so, on-site staff admitted to Phoenix Net Technology: "Progress on the brain is slow, we still have a lot of room for improvement. No one has reached the finish line in one step."
At ZhiPingFang's demo site, each robot was still accompanied by a human operator assisting it in grabbing ice cubes. "It took three to four minutes to make one cup at the beginning; now it can make a simple coffee in about one and a half minutes, but it still can't operate completely automatically without human involvement."
Besides entering factories, the "making milk tea" scenario is also being targeted by more companies. "We just looked around; there are clearly more companies doing things like making coffee or cocktails this year," the above personnel said. Such machines average around 500,000 yuan, with monthly deliveries reaching around a hundred units. "Production line capacity is already struggling to keep up; we're planning for 10,000 machines in the second half of the year."
Behind this judgment is real capital from the market. According to IT桔子 data, in the first half of 2026, total financing in China's embodied intelligence sector reached 93.5 billion yuan, with 14.5 billion flowing into "brain-focused" companies. The market truly understands that the bottleneck of embodied intelligence is not the "body," but the "brain."
Fourier showcased another approach. At this WAIC, Fourier首次公开 "Embodied Home"—a full-stack technology demo for home companionship services. Its core breakthrough lies in no longer needing to tell the robot "go to the kitchen countertop, pick up the cup, put it on the living room table"; just saying "I'm thirsty" allows the robot to autonomously complete the entire process. However, on-site staff also told Phoenix Net Technology that there's still a distance from truly making customers willing to pay and treating it as a usable product. "The brain still can't achieve the effect we expect."
Besides home robots, another product attracting more attention was the desktop robot GR Nano with its distinctive "cat head" design, focusing on emotional value—set for official release in September, priced under a thousand yuan.
LimX Dynamics also joined the "brain" race. On July 15, the company released version 0.5 of its humanoid brain system LimX COSA. Accelerated Evolution globally debuted Booster T2 at this WAIC, positioned as a "next-generation humanoid robot built for future real applications," aiming to push humanoid robots from "able to move" to "usable." Songyan Dynamics brought the Bumi Xiao Bumi OTA V3.0 version. This product, which appeared in this year's Spring Festival Gala and costs around ten thousand yuan, upgraded multiple functions via software OTA updates, enabling automatic connection and newly added discovery functions covering content like Himalaya, children's programming, and the game "Boxing Kid."
The Challenges Before Entering Homes
Mass production is not just a technical issue, but also an economic one.
Data is the first challenge. Embodied intelligence training relies on vast amounts of real-world operational data. The industry's mainstream approach is: letting robots repeatedly execute tasks in simulators, automatically generating massive amounts of "vision + action" paired data at almost zero cost. But this path has its ceiling—the Sim-to-Real gap. Physical laws in simulators always differ from the real world, and trained strategies often fail when transferred to real machines. Simulation is better suited for pre-validation and supplementary enhancement; true generalization ability still relies on real data.
Cost is the second reality. Humanoid robot prices are undergoing a dramatic divergence. Consumer-grade product prices continue to drop—Unitree's dual-arm humanoid robot R1 series released in April 2026 starts at only 26,900 yuan; Songyan Dynamics' Xiao Bumi further lowered the price directly to 9,998 yuan,首次 bringing mass-produced humanoid robots into the sub-10,000-yuan price range. Industrial-grade robot prices remain steady, with wheeled forms becoming the mainstream for industrial application.
ZhiPingFang's AlphaBot 2 averages around 500,000 yuan. Ubtech's launched U1 series full-size ultra-bionic emotional humanoid robots range from 119,800 yuan for the half-body Lite model to 990,000 yuan for the top-tier Ultra model. Founder Zhou Jian坦言 that bionic robot manufacturing processes are still in the exploratory stage, with overall mass production and technical difficulty being "unprecedented in history."
A Goldman Sachs report predicts that reducing the cost of each humanoid robot to 20,000–30,000 USD would significantly promote its adoption across industries.
If the "brain" is the technical bottleneck, then "entering homes" is the unspoken ultimate goal for everyone—also the most distant one.
Regarding the home场景, multiple companies gave almost identical judgments. ZhiPingFang stated clearly: "Entering homes is definitely the goal, but now is not a particularly good时机. First through industrial scenarios, then public services, finally considering homes. Pushing for homes now—globally, no company can truly say they are pushing into homes."
The team from Future Not Far, participating for the first time this year, spun off from the education company Zhangmen three years ago. With home companionship as the场景, they expanded to multiple scenarios like housekeeping services, tidying up, companion interaction, and AI voice interaction. Their launched general-purpose home robot, using a rental model, has entered over 500 households, accumulating 30,000 hours of home working time.
"Besides playing, it can also do basic housework, not replacing tool-type machines, but helping humans complete some preparatory work, like putting clothes into the washing machine or tidying up toys scattered by children." To save costs, it uses grippers instead of dexterous hands.
But on-site staff also admitted that current home-entering robots are more like "large toys" with smart speakers added. Therefore, they target customers who are willing to let their children try new tech, experience cutting-edge results, and have certain payment ability—daily rent 100 yuan, monthly rent 3,000 yuan.
"The renewal rate is quite high, but at our current stage, we hope to reach more families, so we generally only rent for one month." Louis, co-founder of Future Not Far, revealed to Phoenix Net Technology that the robot body price can already be lowered to twenty thousand, but considering computing power costs, the rental model is more reasonable for home-entering robots at this stage.
"The biggest problem with robots now is they can't clean themselves. We only dare to let them do basic tidying and laundry. Once it involves heavy water or greasy stains, it's not possible." Louis predicts that home-entering robots being able to fully replace cleaning阿姨's work "will take at least five more years."
The reason is that the home is a completely unstructured environment with too many uncertainties, plus safety and privacy issues. Even if a company sends a robot into a home, it's more of an experimental exploration, mainly to collect richer home场景 data, still requiring配备 real technicians and even remote teleoperation assistance.
According to MIIT predictions, this year's annual humanoid robot production is expected to exceed 100,000 units. In 2025, global humanoid robot shipments were still around 20,000 units. 2026 is widely recognized by the industry as the "year of mass production" for humanoid robots—not the "year of demos" or "year of financing," but the year of "physical machines rolling off production lines, customers signing receipts, and production lines operating."
The definition of "year of mass production" remains微妙. If it refers to ten-thousand-unit-level shipments, real industrial deployment, and replicable commercial scenarios—then yes, things are indeed happening in 2026 that haven't happened before: Agibot's ten-thousand-unit下线, ZhiPingFang's thousand-unit orders, Unitree's IPO冲刺, Galaxy General's接连 big orders all point in the same direction.
But if "mass production" means entering千家万户 like smartphones, it's still far off. The non-structure, safety, privacy, and cost of home scenarios—each is a hurdle yet to be crossed.
Viewed this way, WAIC 2026 is no longer a stage for humanoid robots to display "flashy tricks," but a "testing ground" for their commercial value. From运动 ability to industrial capability, from technological breakthroughs to commercial落地, from cool displays to stable delivery, humanoid robots are undergoing a rite of passage from "exhibit" to "product."
In the exhibition hall, they seem to have learned how to "work." As for whether they can graduate successfully from vocational school, the answer isn't on the exhibition booths, but in factories, in supermarkets, in those places where they are truly needed to work continuously for 24 hours and repeat actions ten thousand times.
This article is from the WeChat public account: 凤凰网科技 , author: Phoenix Net Technology, editor: Dong Yuqing
